Scenema VAE Encode (voice reference) ⚡

Zero-shot voice cloning has a classic shape: you give the model a snippet of a voice, it encodes that snippet, and the generation is conditioned on the encoded voice instead of a written description. This node is the encode step for Scenema Audio. Feed it the Scenema VAE and a reference clip, and out comes ref_latent - an audio latent that carries the voice's identity, ready to plug into Scenema Audio Generate's ref_latent input.

What it does

It's deliberately small: vae + audio in, ref_latent out. The VAE is the one from Scenema Model Loader (or its vae_encoder_name path). The audio is any loaded clip - Load Audio from core is the obvious source. The node runs the reference through the audio VAE encoder and hands you the latent.

The one knob is max_seconds (default 20, 1–20 in 0.5 steps), and here's the honesty the tooltip gives you: it's hard-capped at 20 seconds. You can't ask for more reference than the encoder's design handles, and you rarely need it - the classic advice for voice references is "as little as you can get away with," and 5–15 seconds of clean, single-speaker audio is the sweet spot. More audio isn't more identity; it's more chance of background noise or a second voice muddying the latent.

How it fits

The chain is: Load Audio → Scenema VAE Encode → Scenema Audio Generate (ref_latent). Once the latent is in, the generate node does zero-shot A2V cloning - it conditions the diffusion generation on your reference voice. The Scenema stack does this properly, too: ref_latent is the LTX A2V-style reference, and there's a separate identity_reference input on Generate for the SeedVC post-pass. Keep them straight: the latent here is the A2V cloning reference; the SeedVC identity reference is a fixed AUDIO you'd feed if you want the final voice-consistency pass locked to a specific voice.

Installing it

Part of the ComfyUI-GGUF-Loader pack under 🤖 CCTech/Scenema. ComfyUI Manager → search "ComfyUI-GGUF-Loader" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
cd ComfyUI-GGUF-Loader
pip install -r requirements.txt

Common issues

Reference quality is the whole game. A clipped, echoey, or multi-speaker reference produces a cloned voice with the same problems - this is the one place in the pipeline where garbage-in is unmistakable. Watch the sample rate and loudness too: if the source clip is at a wildly different level than speech usually sits, normalize it first. And remember the 20s cap - the input won't error on a 40s clip silently; set max_seconds to what you actually want under the cap. If cloning sounds "off" but the reference is clean, try a shorter, more neutral clip - sometimes a few seconds of plain read speech beats a dramatic 15 seconds.